AI Watch: DeepSeek V4 Flash official launch – our initial thought

Research evidence excerpt

Accessible version

Internet - China

AI Watch: DeepSeek V4 Flash official

launch – our initial thought

Industry Overview

Welcome to the latest “AI Watch” series in which we track China’s rapidly evolving AI

landscape and implications for China Internet sector.

DS V4 Flash raises the bar for “capable & cheap” models

DeepSeek (DS) officially launched DeepSeek-V4-Flash-0731 into on July 31. Unlike

traditional model refresh, V4 Flash official version retains the same architecture and

parameter count as the April preview version, with improvements driven entirely by

post-training. DeepSeek highlighted that V4 Flash's agent benchmarks now significantly

exceed those of the earlier V4-Pro Preview despite its smaller model size. On

performance, V4 Flash is among the top global coding and agent-focused models.

Pricing has been one of the key differentiators of DS and V4 Flash is priced unchanged

from preview version, cheaply at US$0.14 / US$0.28 per million input/output tokens, and

substantially below leading Western/Chinese frontier models while approaching

comparable agent benchmark performance. Interestingly, the launch comes amid an

intensifying global AI pricing war as OpenAI cut its GPT-5.6 Luna model pricing by 80%

over the weekend.

Implications to China AI/Internet sector and key stocks

Our high-level takeaway is that V4 Flash demonstrates that post-training and agent

optimization can generate meaningful capability gains without requiring dramatic

parameter scaling. For Alibaba, V4 Flash raises competition intensity for Qwen in

coding and agent benchmarks, although Alibaba should continue benefiting from rising

AI compute demand. For Tencent, the development is broadly neutral as ecosystem

integration and distribution remain the key competitive differentiators. We see the

launch as incrementally negative for Zhipu and MiniMax. With both companies yet to

launch their next flagship upgrades, the burden of proof shifts further toward

demonstrating meaningful performance differentiation. DeepSeek's latest release

reinforces our view that leadership in China's model layer remains highly transient, with

…